Glossary

Every term you'll see on a card, in plain English, with why it matters when you choose or use a model.

Basics

Token: The unit models read and write: a word, part of a word, or punctuation. In English one token is roughly three quarters of a word, but each model's tokenizer counts differently.; Why it matters · Prices and limits are per token, so the same text can cost different amounts on different models.
Context window: The most tokens a model can consider in one request: your prompt, files, conversation history and its own reply together.; Why it matters · Decides whether a whole codebase, contract or long chat fits in one call.
Max output tokens: The longest single reply a model can produce, in tokens. On reasoning models, thinking tokens usually count toward it.; Why it matters · Long generations (full files, reports) need a high limit; set it too low and replies get cut off.
Knowledge cutoff: The date the model's training data is reliable up to. It knows little or nothing about events after it unless you give it the information.; Why it matters · For recent APIs, prices or news, plan to supply the facts or use a search tool.
Latency: How long a request takes: time to the first token and total time. It grows with prompt length, output length and reasoning effort.; Why it matters · Interactive apps need low latency; background jobs can trade it for quality or cost.

Pricing

Input price: What the maker charges per million tokens you send: prompt, documents, images and conversation history.; Why it matters · Dominates the bill for long-document and agent work, where the same context is re-sent many times.
Output price: What the maker charges per million tokens the model generates, including reasoning tokens. Usually 3 to 5 times the input price.; Why it matters · Dominates the bill for generation-heavy work like writing code or long answers.
Prompt caching: Reusing a prompt prefix the provider has already processed. Writing to the cache can cost slightly more than input; reading from it costs a small fraction of input.; Why it matters · Agents and chat apps that re-send the same context can cut input costs by most of the bill.
Batch API: Submitting many requests to be processed asynchronously, with results returned later (often within 24 hours), usually at about half the normal price.; Why it matters · The cheapest way to run evaluations, classification or bulk processing that doesn't need instant answers.
Self-hosting cost: What it costs to run open-weights models on your own or rented hardware: GPUs, serving software, operations and engineering time.; Why it matters · Open weights have no per-token API price, but they are not free to run at scale.

Reasoning

Reasoning tokens: Tokens a model generates while working through a problem before answering. You are billed for them as output, even when they aren't shown.; Why it matters · They improve hard tasks but add cost and latency; budget for them.
Effort setting: An API parameter that controls how much a model reasons: low effort answers faster and cheaper, high effort is slower and more thorough.; Why it matters · Tune it per route: low for chat and classification, high for agents and hard problems.
Adaptive thinking: The model decides how much to reason for each request, guided by the effort setting, instead of you fixing a token budget.; Why it matters · Simple prompts stay fast while hard ones get more thought, without per-request tuning.

Capabilities

Multimodal input: Accepting more than text, such as images, PDFs, audio or video, in the same request.; Why it matters · Needed for screenshots, scanned documents, charts and UI work.
Tool use: Also called function calling. The model asks your code to run a function you described (search, database query, API call) and uses the result.; Why it matters · The basis of agents and of any app where the model takes actions.
Structured outputs: Constraining the model's reply to match a schema, such as a specific JSON shape.; Why it matters · Removes brittle parsing and retry code when other software consumes the output.
Computer use: A model operating a computer through screenshots, mouse and keyboard actions to complete tasks in apps that have no API.; Why it matters · Automates GUI workflows, but needs sandboxing and close supervision.

Openness

Open weights: The trained model parameters are published for download, so you can run, fine-tune and inspect the model yourself, subject to its license.; Why it matters · Gives control over data, hosting and customization; you take on serving and safety work.
Model license: The legal terms for using model weights or outputs: commercial use, redistribution, attribution, user-count limits or regional restrictions.; Why it matters · Some 'open' licenses restrict large companies, certain regions or derivative naming. Read it before shipping.
API model ID: The exact string you pass to an API to select a model, which can differ between the maker's API and cloud platforms.; Why it matters · Pin specific IDs in production so behavior and price don't change under you.

Evaluation

Benchmark: A standardized test set used to score models on a task. Makers often run them with different settings, so scores are only comparable within one report.; Why it matters · Useful for rough comparison; always test on your own tasks before deciding.
SWE-bench: A family of benchmarks where a model must fix real issues in real open-source repositories, checked by the projects' own tests.; Why it matters · One of the better signals for practical coding and agent ability.
Elo rating: A relative score from head-to-head comparisons, where people or graders pick the better of two outputs. Differences matter, not absolute values.; Why it matters · Captures quality on open-ended work that has no single right answer.
Hallucination: When a model states something false with confidence, such as invented facts, citations, APIs or numbers.; Why it matters · Verify claims that matter, and prefer models that say 'I don't know' over ones that guess.

Safety

System card: A maker's report on a model release: capability evaluations, safety and misuse testing, risk determinations and the safeguards it ships with.; Why it matters · The best public evidence of what a model can do and how it can fail.
Model card: A standard document describing a model: intended use, training data, evaluation results, limitations and ethical considerations.; Why it matters · Tells you whether your use case is in scope before you build on the model.
Harmless response rate: In a maker's safety tests, the share of clearly harmful requests where the model's reply did not help cause the harm.; Why it matters · Higher is safer, but compare only within one maker's methodology.
Over-refusal rate: The share of benign but sensitive-sounding requests that a model wrongly refuses.; Why it matters · High over-refusal frustrates users and breaks legitimate products.
Prompt injection: Instructions hidden in content the model reads, such as web pages, emails, files or tool results, that try to override what you or your user asked.; Why it matters · The main security risk for agents with tools; treat all fetched content as untrusted.
Refusal fallback: When a safety classifier blocks a request, the provider re-runs it on a different model instead of returning a refusal.; Why it matters · Keeps legitimate work flowing, but the answer may come from a different model than you chose.
Zero data retention: An arrangement where the provider does not store your prompts or outputs after processing, beyond what's needed to serve the request.; Why it matters · Often required for regulated or sensitive data; not every model or provider offers it.

Media

Price per image: Image models are usually priced per generated image, varying by resolution and quality, or per megapixel.; Why it matters · Compare at the size and quality you will actually use.
Price per second: Video models are usually priced per second of generated video, often more for higher resolution or with sound.; Why it matters · Short clips add up quickly; estimate total seconds per month.
Text-to-speech: Generating spoken audio from text, usually priced per character or per minute of audio.; Why it matters · Voice quality, latency and language coverage vary a lot between models.
Speech-to-text: Transcribing audio into text, usually priced per minute or hour. Diarization labels who spoke when.; Why it matters · Accuracy on your accents, domain words and noisy audio matters more than headline scores.
Content provenance: Signals that mark media as AI-generated, such as C2PA metadata or invisible watermarks like SynthID.; Why it matters · Some platforms and regulations expect it; metadata can be stripped, watermarks are more robust.